A new EM algorithm for underdetermined convolutive blind source separation

This paper presents a new statistical method for separating more than two sound sources from a two-channel recording. It is based on a probabilistic model of the Interchannel Level/Phase Difference presented in [1] and the model parameters are estimated using the maximum likelihood criterion and an Expectation-Maximization algorithm. The source separation task is achieved by soft time-frequency masking of the observation. These masks are derived from the estimated source position model. Algorithm performance is evaluated on the real and synthetic convolutive mixtures data of the first audio source evaluation campaign [2] as well as the Signal Separation Evaluation campaign (SiSEC) [10]. Promising results are obtained when comparing to the other methods presented in these two campaigns.